Calibration of the High‐Resolution Common Land Model in Simulating the Soil Moisture Over the Northeastern China Using an Adaptive Parameter Learning Method

Abstract The growing complexity of land surface models (LSMs) presents significant challenges for parameter calibration. Compared to traditional optimization algorithms, the deep learning‐based optimization framework, namely differentiable parameter learning (dPL), reduces computational costs and achieves greater spatial generalizability. However, the effectiveness of the parameter values derived from dPL in enhancing land surface modeling needs further verification. This study introduced a deep learning‐based adaptive parameter learning (APL) framework for efficiently optimizing key parameters in the Common Land Model (CoLM) to simulate high‐resolution soil moisture (SM) across Northeast China. We began by constructing a surrogate model using long short‐term memory networks to capture the relationships between CoLM parameters, meteorological forcing data, and simulated SM. Initial parameter optimization using the dPL framework improved SM simulations but revealed discrepancies between the performances of surrogate and process‐based models. The APL framework builds upon dPL by iteratively refining surrogate models with expanded training data sets enhancing their ability to approximate the behavior of process‐based models. Evaluations using four metrics—bias, root mean square error, correlation, and Kling–Gupta efficiency—demonstrated that APL outperformed dPL with both frameworks providing robust parameter estimates. This study underscored the potential of deep learning‐based parameter optimization frameworks to overcome traditional calibration challenges in LSMs by improving computational efficiency, enhancing spatial consistency and increasing resilience to uncertainties in forcing and reference data. Finally, we recommended that improving physical coherence in LSMs should not rely solely on adjusting a few parameters but requires a comprehensive approach, including identifying key parameters, employing multiobjective parameter optimization, and, critically, utilizing high‐precision land surface benchmarking data sets.